Predictors of residual disease after debulking surgery in advanced stage ovarian cancer

Abstract

Objective: Optimal debulking with no macroscopic residual disease strongly predicts ovarian cancer survival. The ability to predict likelihood of optimal debulking, which may be partially dependent on tumor biology, could inform clinical decision-making regarding use of neoadjuvant chemotherapy. Thus, we developed a prediction model including epidemiological factors and tumor markers of residual disease after primary debulking surgery.

Methods: Univariate analyses examined associations of 11 pre-diagnosis epidemiologic factors (n=593) and 24 tumor markers (n=204) with debulking status among incident, high-stage, epithelial ovarian cancer cases from the Nurses' Health Studies and New England Case Control study. We used Bayesian model averaging (BMA) to develop prediction models of optimal debulking with 5x5-fold cross-validation and calculated the area under the curve (AUC).

Results: Current aspirin use was associated with lower odds of optimal debulking compared to never use (OR=0.52, 95%CI=0.31-0.86) and two tissue markers, ADRB2 (OR=2.21, 95%CI=1.23-4.41) and FAP (OR=1.91, 95%CI=1.24-3.05) were associated with increased odds of optimal debulking. The BMA selected aspirin, parity, and menopausal status as the epidemiologic/clinical predictors with the posterior effect probability ≥20%. While the prediction model with epidemiologic/clinical predictors had low performance (average AUC=0.49), the model adding tissue biomarkers showed improved, but weak, performance (average AUC=0.62).

Conclusions: Addition of ovarian tumor tissue markers to our multivariable prediction models based on epidemiologic/clinical data slightly improved the model performance, suggesting debulking status may be in part driven by tumor characteristics. Larger studies are warranted to identify those at high risk of poor surgical outcomes informing personalized treatment.

Keywords: debulking; immunohistochemistry; ovarian cancer; prediction model; residual disease; tissue microarray.

Figure 1

Flow chart of the exclusion criteria and study population included in the various prediction modeling.

Figure 2

Tissue marker associations (odds ratio [OR] and 95% confidence intervals [CI]) with optimal debulking status among advanced stage invasive epithelial ovarian cancer cases in NHS/NHSII/NECC (n=166). All models were adjusted for study (NHS/NHSII and NEC). In the logistic regression models 1 = optimally debulked and 0 = sub-optimally debulked.

Figure 3

Prediction modeling of optimal debulking status using epidemiologic and clinical predictors. Average posterior means and associated average posterior probabilities of epidemiologic and clinical predictors being selected in the final prediction model of optimal debulking status and area under the curve (AUC) of the prediction models using Bayesian model averaging among invasive epithelial ovarian cancer cases (n=593) in NHS/NHSII/NEC. We assessed the posterior probability of 11 predictors for possible selection in the final model and conducted five 5-fold cross-validations. The bar chart (A) presents the average posterior mean across the 25 models that were run in total; grey bars denote predictors with average posterior probabilities <20% and yellow bars denote predictors with average posterior probability ≥ 20%. (B) presents the five average AUCs from the five 5-fold cross-validations and in the legend overall AUC, which is the average of the average AUCs from the 5-fold CV, is presented. Type 2 tumors include high-grade serous or poorly differentiated, Transitional/Brenner, Carcinosarcoma, high grade mixed histology.

Figure 4

Prediction modeling of optimal debulking status using tissue markers. Average posterior means and associated average posterior probabilities of tissue markers being selected in the final prediction model of optimal debulking status and area under the curve (AUC) of the prediction models using Bayesian model averaging among invasive epithelial ovarian cancer cases (n=166) in NHS/NHSII/NEC. We assessed the posterior probability of 24 tissue marker predictors for possible selection in the final model and conducted five 5-fold cross-validations. The bar chart (A) presents the average posterior mean across the 25 models that were run in total; grey bars denote predictors with average posterior probabilities <20% and yellow bars denote predictors with average posterior probability ≥ 20%. (B) presents the five average AUCs from the five 5-fold cross-validations and in the legend overall AUC, which is the average of the average AUCs from the 5-fold CV, is presented.

Figure 5

Prediction modeling of optimal debulking status using epidemiologic/clinical predictors and tissue markers. Average posterior means and associated average posterior probabilities of epidemiologic/clinical predictors and tissue markers being selected in the final prediction model of optimal debulking status and area under the curve (AUC) of the prediction models using Bayesian model averaging among invasive epithelial ovarian cancer cases (n=166) in NHS/NHSII/NEC. We assessed the posterior probability of 24 tissue markers in addition to our base model of clinical predict score, which included epidemiologic/clinical variables that had ≥ 20% posterior probability of being selected in the final prediction model of optimal debulking status (i.e. aspirin, parity, and menopausal status), for possible selection in the final model and conducted five 5-fold cross-validations. The bar chart (A) presents the average posterior mean across the 25 models that were run in total; grey bars denote predictors with average posterior probabilities <20% and yellow bars denote predictors with average posterior probability ≥ 20%. (B) presents the five average AUCs from the five 5-fold cross-validations and in the legend overall AUC, which is the average of the average AUCs from the 5-fold CV, is presented.